InstCombineCalls.cpp revision 78f8ef42173a3a9867ed789073d4ddc652fb7ff2
1//===- InstCombineCalls.cpp -----------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visitCall and visitInvoke functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/Support/CallSite.h" 16#include "llvm/Target/TargetData.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/Transforms/Utils/BuildLibCalls.h" 19#include "llvm/Transforms/Utils/Local.h" 20using namespace llvm; 21 22/// getPromotedType - Return the specified type promoted as it would be to pass 23/// though a va_arg area. 24static Type *getPromotedType(Type *Ty) { 25 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 26 if (ITy->getBitWidth() < 32) 27 return Type::getInt32Ty(Ty->getContext()); 28 } 29 return Ty; 30} 31 32 33Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 34 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD); 35 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD); 36 unsigned MinAlign = std::min(DstAlign, SrcAlign); 37 unsigned CopyAlign = MI->getAlignment(); 38 39 if (CopyAlign < MinAlign) { 40 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 41 MinAlign, false)); 42 return MI; 43 } 44 45 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 46 // load/store. 47 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 48 if (MemOpLength == 0) return 0; 49 50 // Source and destination pointer types are always "i8*" for intrinsic. See 51 // if the size is something we can handle with a single primitive load/store. 52 // A single load+store correctly handles overlapping memory in the memmove 53 // case. 54 unsigned Size = MemOpLength->getZExtValue(); 55 if (Size == 0) return MI; // Delete this mem transfer. 56 57 if (Size > 8 || (Size&(Size-1))) 58 return 0; // If not 1/2/4/8 bytes, exit. 59 60 // Use an integer load+store unless we can find something better. 61 unsigned SrcAddrSp = 62 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 63 unsigned DstAddrSp = 64 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 65 66 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 67 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 68 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 69 70 // Memcpy forces the use of i8* for the source and destination. That means 71 // that if you're using memcpy to move one double around, you'll get a cast 72 // from double* to i8*. We'd much rather use a double load+store rather than 73 // an i64 load+store, here because this improves the odds that the source or 74 // dest address will be promotable. See if we can find a better type than the 75 // integer datatype. 76 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 77 if (StrippedDest != MI->getArgOperand(0)) { 78 Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 79 ->getElementType(); 80 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 81 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 82 // down through these levels if so. 83 while (!SrcETy->isSingleValueType()) { 84 if (StructType *STy = dyn_cast<StructType>(SrcETy)) { 85 if (STy->getNumElements() == 1) 86 SrcETy = STy->getElementType(0); 87 else 88 break; 89 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 90 if (ATy->getNumElements() == 1) 91 SrcETy = ATy->getElementType(); 92 else 93 break; 94 } else 95 break; 96 } 97 98 if (SrcETy->isSingleValueType()) { 99 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 100 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 101 } 102 } 103 } 104 105 106 // If the memcpy/memmove provides better alignment info than we can 107 // infer, use it. 108 SrcAlign = std::max(SrcAlign, CopyAlign); 109 DstAlign = std::max(DstAlign, CopyAlign); 110 111 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 112 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 113 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile()); 114 L->setAlignment(SrcAlign); 115 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile()); 116 S->setAlignment(DstAlign); 117 118 // Set the size of the copy to 0, it will be deleted on the next iteration. 119 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 120 return MI; 121} 122 123Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 124 unsigned Alignment = getKnownAlignment(MI->getDest(), TD); 125 if (MI->getAlignment() < Alignment) { 126 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 127 Alignment, false)); 128 return MI; 129 } 130 131 // Extract the length and alignment and fill if they are constant. 132 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 133 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 134 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 135 return 0; 136 uint64_t Len = LenC->getZExtValue(); 137 Alignment = MI->getAlignment(); 138 139 // If the length is zero, this is a no-op 140 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 141 142 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 143 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 144 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 145 146 Value *Dest = MI->getDest(); 147 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 148 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 149 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy); 150 151 // Alignment 0 is identity for alignment 1 for memset, but not store. 152 if (Alignment == 0) Alignment = 1; 153 154 // Extract the fill value and store. 155 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 156 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest, 157 MI->isVolatile()); 158 S->setAlignment(Alignment); 159 160 // Set the size of the copy to 0, it will be deleted on the next iteration. 161 MI->setLength(Constant::getNullValue(LenC->getType())); 162 return MI; 163 } 164 165 return 0; 166} 167 168/// visitCallInst - CallInst simplification. This mostly only handles folding 169/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 170/// the heavy lifting. 171/// 172Instruction *InstCombiner::visitCallInst(CallInst &CI) { 173 if (isFreeCall(&CI)) 174 return visitFree(CI); 175 176 // If the caller function is nounwind, mark the call as nounwind, even if the 177 // callee isn't. 178 if (CI.getParent()->getParent()->doesNotThrow() && 179 !CI.doesNotThrow()) { 180 CI.setDoesNotThrow(); 181 return &CI; 182 } 183 184 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 185 if (!II) return visitCallSite(&CI); 186 187 // Intrinsics cannot occur in an invoke, so handle them here instead of in 188 // visitCallSite. 189 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 190 bool Changed = false; 191 192 // memmove/cpy/set of zero bytes is a noop. 193 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 194 if (NumBytes->isNullValue()) 195 return EraseInstFromFunction(CI); 196 197 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 198 if (CI->getZExtValue() == 1) { 199 // Replace the instruction with just byte operations. We would 200 // transform other cases to loads/stores, but we don't know if 201 // alignment is sufficient. 202 } 203 } 204 205 // No other transformations apply to volatile transfers. 206 if (MI->isVolatile()) 207 return 0; 208 209 // If we have a memmove and the source operation is a constant global, 210 // then the source and dest pointers can't alias, so we can change this 211 // into a call to memcpy. 212 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 213 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 214 if (GVSrc->isConstant()) { 215 Module *M = CI.getParent()->getParent()->getParent(); 216 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 217 Type *Tys[3] = { CI.getArgOperand(0)->getType(), 218 CI.getArgOperand(1)->getType(), 219 CI.getArgOperand(2)->getType() }; 220 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys)); 221 Changed = true; 222 } 223 } 224 225 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 226 // memmove(x,x,size) -> noop. 227 if (MTI->getSource() == MTI->getDest()) 228 return EraseInstFromFunction(CI); 229 } 230 231 // If we can determine a pointer alignment that is bigger than currently 232 // set, update the alignment. 233 if (isa<MemTransferInst>(MI)) { 234 if (Instruction *I = SimplifyMemTransfer(MI)) 235 return I; 236 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 237 if (Instruction *I = SimplifyMemSet(MSI)) 238 return I; 239 } 240 241 if (Changed) return II; 242 } 243 244 switch (II->getIntrinsicID()) { 245 default: break; 246 case Intrinsic::objectsize: { 247 uint64_t Size; 248 if (getObjectSize(II->getArgOperand(0), Size, TD)) 249 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size)); 250 return 0; 251 } 252 case Intrinsic::bswap: 253 // bswap(bswap(x)) -> x 254 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 255 if (Operand->getIntrinsicID() == Intrinsic::bswap) 256 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 257 258 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 259 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 260 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 261 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 262 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 263 TI->getType()->getPrimitiveSizeInBits(); 264 Value *CV = ConstantInt::get(Operand->getType(), C); 265 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 266 return new TruncInst(V, TI->getType()); 267 } 268 } 269 270 break; 271 case Intrinsic::powi: 272 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 273 // powi(x, 0) -> 1.0 274 if (Power->isZero()) 275 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 276 // powi(x, 1) -> x 277 if (Power->isOne()) 278 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 279 // powi(x, -1) -> 1/x 280 if (Power->isAllOnesValue()) 281 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 282 II->getArgOperand(0)); 283 } 284 break; 285 case Intrinsic::cttz: { 286 // If all bits below the first known one are known zero, 287 // this value is constant. 288 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 289 // FIXME: Try to simplify vectors of integers. 290 if (!IT) break; 291 uint32_t BitWidth = IT->getBitWidth(); 292 APInt KnownZero(BitWidth, 0); 293 APInt KnownOne(BitWidth, 0); 294 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 295 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 296 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 297 if ((Mask & KnownZero) == Mask) 298 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 299 APInt(BitWidth, TrailingZeros))); 300 301 } 302 break; 303 case Intrinsic::ctlz: { 304 // If all bits above the first known one are known zero, 305 // this value is constant. 306 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 307 // FIXME: Try to simplify vectors of integers. 308 if (!IT) break; 309 uint32_t BitWidth = IT->getBitWidth(); 310 APInt KnownZero(BitWidth, 0); 311 APInt KnownOne(BitWidth, 0); 312 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 313 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 314 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 315 if ((Mask & KnownZero) == Mask) 316 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 317 APInt(BitWidth, LeadingZeros))); 318 319 } 320 break; 321 case Intrinsic::uadd_with_overflow: { 322 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 323 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 324 uint32_t BitWidth = IT->getBitWidth(); 325 APInt LHSKnownZero(BitWidth, 0); 326 APInt LHSKnownOne(BitWidth, 0); 327 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 328 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 329 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 330 331 if (LHSKnownNegative || LHSKnownPositive) { 332 APInt RHSKnownZero(BitWidth, 0); 333 APInt RHSKnownOne(BitWidth, 0); 334 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 335 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 336 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 337 if (LHSKnownNegative && RHSKnownNegative) { 338 // The sign bit is set in both cases: this MUST overflow. 339 // Create a simple add instruction, and insert it into the struct. 340 Value *Add = Builder->CreateAdd(LHS, RHS); 341 Add->takeName(&CI); 342 Constant *V[] = { 343 UndefValue::get(LHS->getType()), 344 ConstantInt::getTrue(II->getContext()) 345 }; 346 StructType *ST = cast<StructType>(II->getType()); 347 Constant *Struct = ConstantStruct::get(ST, V); 348 return InsertValueInst::Create(Struct, Add, 0); 349 } 350 351 if (LHSKnownPositive && RHSKnownPositive) { 352 // The sign bit is clear in both cases: this CANNOT overflow. 353 // Create a simple add instruction, and insert it into the struct. 354 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 355 Add->takeName(&CI); 356 Constant *V[] = { 357 UndefValue::get(LHS->getType()), 358 ConstantInt::getFalse(II->getContext()) 359 }; 360 StructType *ST = cast<StructType>(II->getType()); 361 Constant *Struct = ConstantStruct::get(ST, V); 362 return InsertValueInst::Create(Struct, Add, 0); 363 } 364 } 365 } 366 // FALL THROUGH uadd into sadd 367 case Intrinsic::sadd_with_overflow: 368 // Canonicalize constants into the RHS. 369 if (isa<Constant>(II->getArgOperand(0)) && 370 !isa<Constant>(II->getArgOperand(1))) { 371 Value *LHS = II->getArgOperand(0); 372 II->setArgOperand(0, II->getArgOperand(1)); 373 II->setArgOperand(1, LHS); 374 return II; 375 } 376 377 // X + undef -> undef 378 if (isa<UndefValue>(II->getArgOperand(1))) 379 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 380 381 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 382 // X + 0 -> {X, false} 383 if (RHS->isZero()) { 384 Constant *V[] = { 385 UndefValue::get(II->getArgOperand(0)->getType()), 386 ConstantInt::getFalse(II->getContext()) 387 }; 388 Constant *Struct = 389 ConstantStruct::get(cast<StructType>(II->getType()), V); 390 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 391 } 392 } 393 break; 394 case Intrinsic::usub_with_overflow: 395 case Intrinsic::ssub_with_overflow: 396 // undef - X -> undef 397 // X - undef -> undef 398 if (isa<UndefValue>(II->getArgOperand(0)) || 399 isa<UndefValue>(II->getArgOperand(1))) 400 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 401 402 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 403 // X - 0 -> {X, false} 404 if (RHS->isZero()) { 405 Constant *V[] = { 406 UndefValue::get(II->getArgOperand(0)->getType()), 407 ConstantInt::getFalse(II->getContext()) 408 }; 409 Constant *Struct = 410 ConstantStruct::get(cast<StructType>(II->getType()), V); 411 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 412 } 413 } 414 break; 415 case Intrinsic::umul_with_overflow: { 416 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 417 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 418 419 APInt LHSKnownZero(BitWidth, 0); 420 APInt LHSKnownOne(BitWidth, 0); 421 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 422 APInt RHSKnownZero(BitWidth, 0); 423 APInt RHSKnownOne(BitWidth, 0); 424 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 425 426 // Get the largest possible values for each operand. 427 APInt LHSMax = ~LHSKnownZero; 428 APInt RHSMax = ~RHSKnownZero; 429 430 // If multiplying the maximum values does not overflow then we can turn 431 // this into a plain NUW mul. 432 bool Overflow; 433 LHSMax.umul_ov(RHSMax, Overflow); 434 if (!Overflow) { 435 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 436 Constant *V[] = { 437 UndefValue::get(LHS->getType()), 438 Builder->getFalse() 439 }; 440 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 441 return InsertValueInst::Create(Struct, Mul, 0); 442 } 443 } // FALL THROUGH 444 case Intrinsic::smul_with_overflow: 445 // Canonicalize constants into the RHS. 446 if (isa<Constant>(II->getArgOperand(0)) && 447 !isa<Constant>(II->getArgOperand(1))) { 448 Value *LHS = II->getArgOperand(0); 449 II->setArgOperand(0, II->getArgOperand(1)); 450 II->setArgOperand(1, LHS); 451 return II; 452 } 453 454 // X * undef -> undef 455 if (isa<UndefValue>(II->getArgOperand(1))) 456 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 457 458 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 459 // X*0 -> {0, false} 460 if (RHSI->isZero()) 461 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 462 463 // X * 1 -> {X, false} 464 if (RHSI->equalsInt(1)) { 465 Constant *V[] = { 466 UndefValue::get(II->getArgOperand(0)->getType()), 467 ConstantInt::getFalse(II->getContext()) 468 }; 469 Constant *Struct = 470 ConstantStruct::get(cast<StructType>(II->getType()), V); 471 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 472 } 473 } 474 break; 475 case Intrinsic::ppc_altivec_lvx: 476 case Intrinsic::ppc_altivec_lvxl: 477 // Turn PPC lvx -> load if the pointer is known aligned. 478 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 479 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 480 PointerType::getUnqual(II->getType())); 481 return new LoadInst(Ptr); 482 } 483 break; 484 case Intrinsic::ppc_altivec_stvx: 485 case Intrinsic::ppc_altivec_stvxl: 486 // Turn stvx -> store if the pointer is known aligned. 487 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 488 Type *OpPtrTy = 489 PointerType::getUnqual(II->getArgOperand(0)->getType()); 490 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 491 return new StoreInst(II->getArgOperand(0), Ptr); 492 } 493 break; 494 case Intrinsic::x86_sse_storeu_ps: 495 case Intrinsic::x86_sse2_storeu_pd: 496 case Intrinsic::x86_sse2_storeu_dq: 497 // Turn X86 storeu -> store if the pointer is known aligned. 498 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 499 Type *OpPtrTy = 500 PointerType::getUnqual(II->getArgOperand(1)->getType()); 501 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 502 return new StoreInst(II->getArgOperand(1), Ptr); 503 } 504 break; 505 506 case Intrinsic::x86_sse_cvtss2si: 507 case Intrinsic::x86_sse_cvtss2si64: 508 case Intrinsic::x86_sse_cvttss2si: 509 case Intrinsic::x86_sse_cvttss2si64: 510 case Intrinsic::x86_sse2_cvtsd2si: 511 case Intrinsic::x86_sse2_cvtsd2si64: 512 case Intrinsic::x86_sse2_cvttsd2si: 513 case Intrinsic::x86_sse2_cvttsd2si64: { 514 // These intrinsics only demand the 0th element of their input vectors. If 515 // we can simplify the input based on that, do so now. 516 unsigned VWidth = 517 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 518 APInt DemandedElts(VWidth, 1); 519 APInt UndefElts(VWidth, 0); 520 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 521 DemandedElts, UndefElts)) { 522 II->setArgOperand(0, V); 523 return II; 524 } 525 break; 526 } 527 528 529 case Intrinsic::x86_sse41_pmovsxbw: 530 case Intrinsic::x86_sse41_pmovsxwd: 531 case Intrinsic::x86_sse41_pmovsxdq: 532 case Intrinsic::x86_sse41_pmovzxbw: 533 case Intrinsic::x86_sse41_pmovzxwd: 534 case Intrinsic::x86_sse41_pmovzxdq: { 535 // pmov{s|z}x ignores the upper half of their input vectors. 536 unsigned VWidth = 537 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 538 unsigned LowHalfElts = VWidth / 2; 539 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 540 APInt UndefElts(VWidth, 0); 541 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 542 InputDemandedElts, 543 UndefElts)) { 544 II->setArgOperand(0, TmpV); 545 return II; 546 } 547 break; 548 } 549 550 case Intrinsic::ppc_altivec_vperm: 551 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 552 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 553 assert(Mask->getType()->getVectorNumElements() == 16 && 554 "Bad type for intrinsic!"); 555 556 // Check that all of the elements are integer constants or undefs. 557 bool AllEltsOk = true; 558 for (unsigned i = 0; i != 16; ++i) { 559 Constant *Elt = Mask->getAggregateElement(i); 560 if (Elt == 0 || 561 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 562 AllEltsOk = false; 563 break; 564 } 565 } 566 567 if (AllEltsOk) { 568 // Cast the input vectors to byte vectors. 569 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 570 Mask->getType()); 571 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 572 Mask->getType()); 573 Value *Result = UndefValue::get(Op0->getType()); 574 575 // Only extract each element once. 576 Value *ExtractedElts[32]; 577 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 578 579 for (unsigned i = 0; i != 16; ++i) { 580 if (isa<UndefValue>(Mask->getAggregateElement(i))) 581 continue; 582 unsigned Idx = 583 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 584 Idx &= 31; // Match the hardware behavior. 585 586 if (ExtractedElts[Idx] == 0) { 587 ExtractedElts[Idx] = 588 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 589 Builder->getInt32(Idx&15)); 590 } 591 592 // Insert this value into the result vector. 593 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 594 Builder->getInt32(i)); 595 } 596 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 597 } 598 } 599 break; 600 601 case Intrinsic::arm_neon_vld1: 602 case Intrinsic::arm_neon_vld2: 603 case Intrinsic::arm_neon_vld3: 604 case Intrinsic::arm_neon_vld4: 605 case Intrinsic::arm_neon_vld2lane: 606 case Intrinsic::arm_neon_vld3lane: 607 case Intrinsic::arm_neon_vld4lane: 608 case Intrinsic::arm_neon_vst1: 609 case Intrinsic::arm_neon_vst2: 610 case Intrinsic::arm_neon_vst3: 611 case Intrinsic::arm_neon_vst4: 612 case Intrinsic::arm_neon_vst2lane: 613 case Intrinsic::arm_neon_vst3lane: 614 case Intrinsic::arm_neon_vst4lane: { 615 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 616 unsigned AlignArg = II->getNumArgOperands() - 1; 617 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 618 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 619 II->setArgOperand(AlignArg, 620 ConstantInt::get(Type::getInt32Ty(II->getContext()), 621 MemAlign, false)); 622 return II; 623 } 624 break; 625 } 626 627 case Intrinsic::arm_neon_vmulls: 628 case Intrinsic::arm_neon_vmullu: { 629 Value *Arg0 = II->getArgOperand(0); 630 Value *Arg1 = II->getArgOperand(1); 631 632 // Handle mul by zero first: 633 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 634 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 635 } 636 637 // Check for constant LHS & RHS - in this case we just simplify. 638 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu); 639 VectorType *NewVT = cast<VectorType>(II->getType()); 640 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth(); 641 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) { 642 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 643 VectorType* VT = cast<VectorType>(CV0->getType()); 644 SmallVector<Constant*, 4> NewElems; 645 for (unsigned i = 0; i < VT->getNumElements(); ++i) { 646 APInt CV0E = 647 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue(); 648 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth); 649 APInt CV1E = 650 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue(); 651 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth); 652 NewElems.push_back( 653 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E)); 654 } 655 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems)); 656 } 657 658 // Couldn't simplify - cannonicalize constant to the RHS. 659 std::swap(Arg0, Arg1); 660 } 661 662 // Handle mul by one: 663 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 664 if (ConstantInt *Splat = 665 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) { 666 if (Splat->isOne()) { 667 if (Zext) 668 return CastInst::CreateZExtOrBitCast(Arg0, II->getType()); 669 // else 670 return CastInst::CreateSExtOrBitCast(Arg0, II->getType()); 671 } 672 } 673 } 674 675 break; 676 } 677 678 case Intrinsic::stackrestore: { 679 // If the save is right next to the restore, remove the restore. This can 680 // happen when variable allocas are DCE'd. 681 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 682 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 683 BasicBlock::iterator BI = SS; 684 if (&*++BI == II) 685 return EraseInstFromFunction(CI); 686 } 687 } 688 689 // Scan down this block to see if there is another stack restore in the 690 // same block without an intervening call/alloca. 691 BasicBlock::iterator BI = II; 692 TerminatorInst *TI = II->getParent()->getTerminator(); 693 bool CannotRemove = false; 694 for (++BI; &*BI != TI; ++BI) { 695 if (isa<AllocaInst>(BI)) { 696 CannotRemove = true; 697 break; 698 } 699 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 700 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 701 // If there is a stackrestore below this one, remove this one. 702 if (II->getIntrinsicID() == Intrinsic::stackrestore) 703 return EraseInstFromFunction(CI); 704 // Otherwise, ignore the intrinsic. 705 } else { 706 // If we found a non-intrinsic call, we can't remove the stack 707 // restore. 708 CannotRemove = true; 709 break; 710 } 711 } 712 } 713 714 // If the stack restore is in a return, resume, or unwind block and if there 715 // are no allocas or calls between the restore and the return, nuke the 716 // restore. 717 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 718 return EraseInstFromFunction(CI); 719 break; 720 } 721 } 722 723 return visitCallSite(II); 724} 725 726// InvokeInst simplification 727// 728Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 729 return visitCallSite(&II); 730} 731 732/// isSafeToEliminateVarargsCast - If this cast does not affect the value 733/// passed through the varargs area, we can eliminate the use of the cast. 734static bool isSafeToEliminateVarargsCast(const CallSite CS, 735 const CastInst * const CI, 736 const TargetData * const TD, 737 const int ix) { 738 if (!CI->isLosslessCast()) 739 return false; 740 741 // The size of ByVal arguments is derived from the type, so we 742 // can't change to a type with a different size. If the size were 743 // passed explicitly we could avoid this check. 744 if (!CS.isByValArgument(ix)) 745 return true; 746 747 Type* SrcTy = 748 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 749 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 750 if (!SrcTy->isSized() || !DstTy->isSized()) 751 return false; 752 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 753 return false; 754 return true; 755} 756 757namespace { 758class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 759 InstCombiner *IC; 760protected: 761 void replaceCall(Value *With) { 762 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 763 } 764 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 765 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 766 return true; 767 if (ConstantInt *SizeCI = 768 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 769 if (SizeCI->isAllOnesValue()) 770 return true; 771 if (isString) { 772 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 773 // If the length is 0 we don't know how long it is and so we can't 774 // remove the check. 775 if (Len == 0) return false; 776 return SizeCI->getZExtValue() >= Len; 777 } 778 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 779 CI->getArgOperand(SizeArgOp))) 780 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 781 } 782 return false; 783 } 784public: 785 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 786 Instruction *NewInstruction; 787}; 788} // end anonymous namespace 789 790// Try to fold some different type of calls here. 791// Currently we're only working with the checking functions, memcpy_chk, 792// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 793// strcat_chk and strncat_chk. 794Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 795 if (CI->getCalledFunction() == 0) return 0; 796 797 InstCombineFortifiedLibCalls Simplifier(this); 798 Simplifier.fold(CI, TD); 799 return Simplifier.NewInstruction; 800} 801 802static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 803 // Strip off at most one level of pointer casts, looking for an alloca. This 804 // is good enough in practice and simpler than handling any number of casts. 805 Value *Underlying = TrampMem->stripPointerCasts(); 806 if (Underlying != TrampMem && 807 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 808 return 0; 809 if (!isa<AllocaInst>(Underlying)) 810 return 0; 811 812 IntrinsicInst *InitTrampoline = 0; 813 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 814 I != E; I++) { 815 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 816 if (!II) 817 return 0; 818 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 819 if (InitTrampoline) 820 // More than one init_trampoline writes to this value. Give up. 821 return 0; 822 InitTrampoline = II; 823 continue; 824 } 825 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 826 // Allow any number of calls to adjust.trampoline. 827 continue; 828 return 0; 829 } 830 831 // No call to init.trampoline found. 832 if (!InitTrampoline) 833 return 0; 834 835 // Check that the alloca is being used in the expected way. 836 if (InitTrampoline->getOperand(0) != TrampMem) 837 return 0; 838 839 return InitTrampoline; 840} 841 842static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 843 Value *TrampMem) { 844 // Visit all the previous instructions in the basic block, and try to find a 845 // init.trampoline which has a direct path to the adjust.trampoline. 846 for (BasicBlock::iterator I = AdjustTramp, 847 E = AdjustTramp->getParent()->begin(); I != E; ) { 848 Instruction *Inst = --I; 849 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 850 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 851 II->getOperand(0) == TrampMem) 852 return II; 853 if (Inst->mayWriteToMemory()) 854 return 0; 855 } 856 return 0; 857} 858 859// Given a call to llvm.adjust.trampoline, find and return the corresponding 860// call to llvm.init.trampoline if the call to the trampoline can be optimized 861// to a direct call to a function. Otherwise return NULL. 862// 863static IntrinsicInst *FindInitTrampoline(Value *Callee) { 864 Callee = Callee->stripPointerCasts(); 865 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 866 if (!AdjustTramp || 867 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 868 return 0; 869 870 Value *TrampMem = AdjustTramp->getOperand(0); 871 872 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 873 return IT; 874 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 875 return IT; 876 return 0; 877} 878 879// visitCallSite - Improvements for call and invoke instructions. 880// 881Instruction *InstCombiner::visitCallSite(CallSite CS) { 882 if (isAllocLikeFn(CS.getInstruction())) 883 return visitAllocSite(*CS.getInstruction()); 884 885 bool Changed = false; 886 887 // If the callee is a pointer to a function, attempt to move any casts to the 888 // arguments of the call/invoke. 889 Value *Callee = CS.getCalledValue(); 890 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 891 return 0; 892 893 if (Function *CalleeF = dyn_cast<Function>(Callee)) 894 // If the call and callee calling conventions don't match, this call must 895 // be unreachable, as the call is undefined. 896 if (CalleeF->getCallingConv() != CS.getCallingConv() && 897 // Only do this for calls to a function with a body. A prototype may 898 // not actually end up matching the implementation's calling conv for a 899 // variety of reasons (e.g. it may be written in assembly). 900 !CalleeF->isDeclaration()) { 901 Instruction *OldCall = CS.getInstruction(); 902 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 903 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 904 OldCall); 905 // If OldCall dues not return void then replaceAllUsesWith undef. 906 // This allows ValueHandlers and custom metadata to adjust itself. 907 if (!OldCall->getType()->isVoidTy()) 908 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 909 if (isa<CallInst>(OldCall)) 910 return EraseInstFromFunction(*OldCall); 911 912 // We cannot remove an invoke, because it would change the CFG, just 913 // change the callee to a null pointer. 914 cast<InvokeInst>(OldCall)->setCalledFunction( 915 Constant::getNullValue(CalleeF->getType())); 916 return 0; 917 } 918 919 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 920 // If CS does not return void then replaceAllUsesWith undef. 921 // This allows ValueHandlers and custom metadata to adjust itself. 922 if (!CS.getInstruction()->getType()->isVoidTy()) 923 ReplaceInstUsesWith(*CS.getInstruction(), 924 UndefValue::get(CS.getInstruction()->getType())); 925 926 if (isa<InvokeInst>(CS.getInstruction())) { 927 // Can't remove an invoke because we cannot change the CFG. 928 return 0; 929 } 930 931 // This instruction is not reachable, just remove it. We insert a store to 932 // undef so that we know that this code is not reachable, despite the fact 933 // that we can't modify the CFG here. 934 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 935 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 936 CS.getInstruction()); 937 938 return EraseInstFromFunction(*CS.getInstruction()); 939 } 940 941 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 942 return transformCallThroughTrampoline(CS, II); 943 944 PointerType *PTy = cast<PointerType>(Callee->getType()); 945 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 946 if (FTy->isVarArg()) { 947 int ix = FTy->getNumParams(); 948 // See if we can optimize any arguments passed through the varargs area of 949 // the call. 950 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 951 E = CS.arg_end(); I != E; ++I, ++ix) { 952 CastInst *CI = dyn_cast<CastInst>(*I); 953 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 954 *I = CI->getOperand(0); 955 Changed = true; 956 } 957 } 958 } 959 960 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 961 // Inline asm calls cannot throw - mark them 'nounwind'. 962 CS.setDoesNotThrow(); 963 Changed = true; 964 } 965 966 // Try to optimize the call if possible, we require TargetData for most of 967 // this. None of these calls are seen as possibly dead so go ahead and 968 // delete the instruction now. 969 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 970 Instruction *I = tryOptimizeCall(CI, TD); 971 // If we changed something return the result, etc. Otherwise let 972 // the fallthrough check. 973 if (I) return EraseInstFromFunction(*I); 974 } 975 976 return Changed ? CS.getInstruction() : 0; 977} 978 979// transformConstExprCastCall - If the callee is a constexpr cast of a function, 980// attempt to move the cast to the arguments of the call/invoke. 981// 982bool InstCombiner::transformConstExprCastCall(CallSite CS) { 983 Function *Callee = 984 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 985 if (Callee == 0) 986 return false; 987 Instruction *Caller = CS.getInstruction(); 988 const AttrListPtr &CallerPAL = CS.getAttributes(); 989 990 // Okay, this is a cast from a function to a different type. Unless doing so 991 // would cause a type conversion of one of our arguments, change this call to 992 // be a direct call with arguments casted to the appropriate types. 993 // 994 FunctionType *FT = Callee->getFunctionType(); 995 Type *OldRetTy = Caller->getType(); 996 Type *NewRetTy = FT->getReturnType(); 997 998 if (NewRetTy->isStructTy()) 999 return false; // TODO: Handle multiple return values. 1000 1001 // Check to see if we are changing the return type... 1002 if (OldRetTy != NewRetTy) { 1003 if (Callee->isDeclaration() && 1004 // Conversion is ok if changing from one pointer type to another or from 1005 // a pointer to an integer of the same size. 1006 !((OldRetTy->isPointerTy() || !TD || 1007 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1008 (NewRetTy->isPointerTy() || !TD || 1009 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1010 return false; // Cannot transform this return value. 1011 1012 if (!Caller->use_empty() && 1013 // void -> non-void is handled specially 1014 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1015 return false; // Cannot transform this return value. 1016 1017 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1018 Attributes RAttrs = CallerPAL.getRetAttributes(); 1019 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1020 return false; // Attribute not compatible with transformed value. 1021 } 1022 1023 // If the callsite is an invoke instruction, and the return value is used by 1024 // a PHI node in a successor, we cannot change the return type of the call 1025 // because there is no place to put the cast instruction (without breaking 1026 // the critical edge). Bail out in this case. 1027 if (!Caller->use_empty()) 1028 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1029 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1030 UI != E; ++UI) 1031 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1032 if (PN->getParent() == II->getNormalDest() || 1033 PN->getParent() == II->getUnwindDest()) 1034 return false; 1035 } 1036 1037 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1038 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1039 1040 CallSite::arg_iterator AI = CS.arg_begin(); 1041 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1042 Type *ParamTy = FT->getParamType(i); 1043 Type *ActTy = (*AI)->getType(); 1044 1045 if (!CastInst::isCastable(ActTy, ParamTy)) 1046 return false; // Cannot transform this parameter value. 1047 1048 Attributes Attrs = CallerPAL.getParamAttributes(i + 1); 1049 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1050 return false; // Attribute not compatible with transformed value. 1051 1052 // If the parameter is passed as a byval argument, then we have to have a 1053 // sized type and the sized type has to have the same size as the old type. 1054 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1055 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1056 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1057 return false; 1058 1059 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1060 if (TD->getTypeAllocSize(CurElTy) != 1061 TD->getTypeAllocSize(ParamPTy->getElementType())) 1062 return false; 1063 } 1064 1065 // Converting from one pointer type to another or between a pointer and an 1066 // integer of the same size is safe even if we do not have a body. 1067 bool isConvertible = ActTy == ParamTy || 1068 (TD && ((ParamTy->isPointerTy() || 1069 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1070 (ActTy->isPointerTy() || 1071 ActTy == TD->getIntPtrType(Caller->getContext())))); 1072 if (Callee->isDeclaration() && !isConvertible) return false; 1073 } 1074 1075 if (Callee->isDeclaration()) { 1076 // Do not delete arguments unless we have a function body. 1077 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1078 return false; 1079 1080 // If the callee is just a declaration, don't change the varargsness of the 1081 // call. We don't want to introduce a varargs call where one doesn't 1082 // already exist. 1083 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1084 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1085 return false; 1086 1087 // If both the callee and the cast type are varargs, we still have to make 1088 // sure the number of fixed parameters are the same or we have the same 1089 // ABI issues as if we introduce a varargs call. 1090 if (FT->isVarArg() && 1091 cast<FunctionType>(APTy->getElementType())->isVarArg() && 1092 FT->getNumParams() != 1093 cast<FunctionType>(APTy->getElementType())->getNumParams()) 1094 return false; 1095 } 1096 1097 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1098 !CallerPAL.isEmpty()) 1099 // In this case we have more arguments than the new function type, but we 1100 // won't be dropping them. Check that these extra arguments have attributes 1101 // that are compatible with being a vararg call argument. 1102 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1103 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1104 break; 1105 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1106 if (PAttrs & Attribute::VarArgsIncompatible) 1107 return false; 1108 } 1109 1110 1111 // Okay, we decided that this is a safe thing to do: go ahead and start 1112 // inserting cast instructions as necessary. 1113 std::vector<Value*> Args; 1114 Args.reserve(NumActualArgs); 1115 SmallVector<AttributeWithIndex, 8> attrVec; 1116 attrVec.reserve(NumCommonArgs); 1117 1118 // Get any return attributes. 1119 Attributes RAttrs = CallerPAL.getRetAttributes(); 1120 1121 // If the return value is not being used, the type may not be compatible 1122 // with the existing attributes. Wipe out any problematic attributes. 1123 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1124 1125 // Add the new return attributes. 1126 if (RAttrs) 1127 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1128 1129 AI = CS.arg_begin(); 1130 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1131 Type *ParamTy = FT->getParamType(i); 1132 if ((*AI)->getType() == ParamTy) { 1133 Args.push_back(*AI); 1134 } else { 1135 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1136 false, ParamTy, false); 1137 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy)); 1138 } 1139 1140 // Add any parameter attributes. 1141 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1142 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1143 } 1144 1145 // If the function takes more arguments than the call was taking, add them 1146 // now. 1147 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1148 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1149 1150 // If we are removing arguments to the function, emit an obnoxious warning. 1151 if (FT->getNumParams() < NumActualArgs) { 1152 if (!FT->isVarArg()) { 1153 errs() << "WARNING: While resolving call to function '" 1154 << Callee->getName() << "' arguments were dropped!\n"; 1155 } else { 1156 // Add all of the arguments in their promoted form to the arg list. 1157 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1158 Type *PTy = getPromotedType((*AI)->getType()); 1159 if (PTy != (*AI)->getType()) { 1160 // Must promote to pass through va_arg area! 1161 Instruction::CastOps opcode = 1162 CastInst::getCastOpcode(*AI, false, PTy, false); 1163 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 1164 } else { 1165 Args.push_back(*AI); 1166 } 1167 1168 // Add any parameter attributes. 1169 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1170 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1171 } 1172 } 1173 } 1174 1175 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1176 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1177 1178 if (NewRetTy->isVoidTy()) 1179 Caller->setName(""); // Void type should not have a name. 1180 1181 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec); 1182 1183 Instruction *NC; 1184 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1185 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1186 II->getUnwindDest(), Args); 1187 NC->takeName(II); 1188 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1189 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1190 } else { 1191 CallInst *CI = cast<CallInst>(Caller); 1192 NC = Builder->CreateCall(Callee, Args); 1193 NC->takeName(CI); 1194 if (CI->isTailCall()) 1195 cast<CallInst>(NC)->setTailCall(); 1196 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1197 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1198 } 1199 1200 // Insert a cast of the return type as necessary. 1201 Value *NV = NC; 1202 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1203 if (!NV->getType()->isVoidTy()) { 1204 Instruction::CastOps opcode = 1205 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1206 NV = NC = CastInst::Create(opcode, NC, OldRetTy); 1207 NC->setDebugLoc(Caller->getDebugLoc()); 1208 1209 // If this is an invoke instruction, we should insert it after the first 1210 // non-phi, instruction in the normal successor block. 1211 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1212 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1213 InsertNewInstBefore(NC, *I); 1214 } else { 1215 // Otherwise, it's a call, just insert cast right after the call. 1216 InsertNewInstBefore(NC, *Caller); 1217 } 1218 Worklist.AddUsersToWorkList(*Caller); 1219 } else { 1220 NV = UndefValue::get(Caller->getType()); 1221 } 1222 } 1223 1224 if (!Caller->use_empty()) 1225 ReplaceInstUsesWith(*Caller, NV); 1226 1227 EraseInstFromFunction(*Caller); 1228 return true; 1229} 1230 1231// transformCallThroughTrampoline - Turn a call to a function created by 1232// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1233// underlying function. 1234// 1235Instruction * 1236InstCombiner::transformCallThroughTrampoline(CallSite CS, 1237 IntrinsicInst *Tramp) { 1238 Value *Callee = CS.getCalledValue(); 1239 PointerType *PTy = cast<PointerType>(Callee->getType()); 1240 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1241 const AttrListPtr &Attrs = CS.getAttributes(); 1242 1243 // If the call already has the 'nest' attribute somewhere then give up - 1244 // otherwise 'nest' would occur twice after splicing in the chain. 1245 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1246 return 0; 1247 1248 assert(Tramp && 1249 "transformCallThroughTrampoline called with incorrect CallSite."); 1250 1251 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1252 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1253 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1254 1255 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1256 if (!NestAttrs.isEmpty()) { 1257 unsigned NestIdx = 1; 1258 Type *NestTy = 0; 1259 Attributes NestAttr = Attribute::None; 1260 1261 // Look for a parameter marked with the 'nest' attribute. 1262 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1263 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1264 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1265 // Record the parameter type and any other attributes. 1266 NestTy = *I; 1267 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1268 break; 1269 } 1270 1271 if (NestTy) { 1272 Instruction *Caller = CS.getInstruction(); 1273 std::vector<Value*> NewArgs; 1274 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1275 1276 SmallVector<AttributeWithIndex, 8> NewAttrs; 1277 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1278 1279 // Insert the nest argument into the call argument list, which may 1280 // mean appending it. Likewise for attributes. 1281 1282 // Add any result attributes. 1283 if (Attributes Attr = Attrs.getRetAttributes()) 1284 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1285 1286 { 1287 unsigned Idx = 1; 1288 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1289 do { 1290 if (Idx == NestIdx) { 1291 // Add the chain argument and attributes. 1292 Value *NestVal = Tramp->getArgOperand(2); 1293 if (NestVal->getType() != NestTy) 1294 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1295 NewArgs.push_back(NestVal); 1296 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1297 } 1298 1299 if (I == E) 1300 break; 1301 1302 // Add the original argument and attributes. 1303 NewArgs.push_back(*I); 1304 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1305 NewAttrs.push_back 1306 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1307 1308 ++Idx, ++I; 1309 } while (1); 1310 } 1311 1312 // Add any function attributes. 1313 if (Attributes Attr = Attrs.getFnAttributes()) 1314 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1315 1316 // The trampoline may have been bitcast to a bogus type (FTy). 1317 // Handle this by synthesizing a new function type, equal to FTy 1318 // with the chain parameter inserted. 1319 1320 std::vector<Type*> NewTypes; 1321 NewTypes.reserve(FTy->getNumParams()+1); 1322 1323 // Insert the chain's type into the list of parameter types, which may 1324 // mean appending it. 1325 { 1326 unsigned Idx = 1; 1327 FunctionType::param_iterator I = FTy->param_begin(), 1328 E = FTy->param_end(); 1329 1330 do { 1331 if (Idx == NestIdx) 1332 // Add the chain's type. 1333 NewTypes.push_back(NestTy); 1334 1335 if (I == E) 1336 break; 1337 1338 // Add the original type. 1339 NewTypes.push_back(*I); 1340 1341 ++Idx, ++I; 1342 } while (1); 1343 } 1344 1345 // Replace the trampoline call with a direct call. Let the generic 1346 // code sort out any function type mismatches. 1347 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1348 FTy->isVarArg()); 1349 Constant *NewCallee = 1350 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1351 NestF : ConstantExpr::getBitCast(NestF, 1352 PointerType::getUnqual(NewFTy)); 1353 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs); 1354 1355 Instruction *NewCaller; 1356 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1357 NewCaller = InvokeInst::Create(NewCallee, 1358 II->getNormalDest(), II->getUnwindDest(), 1359 NewArgs); 1360 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1361 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1362 } else { 1363 NewCaller = CallInst::Create(NewCallee, NewArgs); 1364 if (cast<CallInst>(Caller)->isTailCall()) 1365 cast<CallInst>(NewCaller)->setTailCall(); 1366 cast<CallInst>(NewCaller)-> 1367 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1368 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1369 } 1370 1371 return NewCaller; 1372 } 1373 } 1374 1375 // Replace the trampoline call with a direct call. Since there is no 'nest' 1376 // parameter, there is no need to adjust the argument list. Let the generic 1377 // code sort out any function type mismatches. 1378 Constant *NewCallee = 1379 NestF->getType() == PTy ? NestF : 1380 ConstantExpr::getBitCast(NestF, PTy); 1381 CS.setCalledFunction(NewCallee); 1382 return CS.getInstruction(); 1383} 1384